Actual-time dental picture verification with Amazon SageMaker AI at Henry Schein One

In dentistry, picture high quality determines whether or not a declare is paid or denied. As much as 20 % insurance coverage claims are initially denied, with lacking or low-quality photos among the many main causes. But high quality evaluation has historically been a guide, after-the-fact course of. A clinician opinions an X-ray hours or days after seize, discovering issues solely when a declare is rejected or a therapy plan can’t proceed. If the picture is blurry, misaligned, or incomplete, the affected person should return for a retake, including value, delay, and frustration for everybody concerned. The basic hole is timing: high quality suggestions arrives lengthy after the affected person has left and the medical second has handed.

This submit describes how Henry Schein One closed that hole by constructing Picture Confirm, an AI-powered high quality verification system on Amazon SageMaker AI that evaluates dental X-ray high quality on the level of seize, in actual time, throughout hundreds of areas. The system went from idea to over 10,000 lively areas inside months and has already processed over 11 million X-rays and rising at 1.5 million per week. Henry Schein One is now scaling towards 40,000 areas globally throughout 4 areas.

The problem: Actual-time picture high quality at scale

Henry Schein One’s earlier picture verification resolution ran on a distinct cloud platform, nevertheless it couldn’t ship the latency or value effectivity required for a easy medical workflow. Rebuilding on AWS wasn’t a migration. It required designing a system that might meet 5 simultaneous necessities. Lacking any one in every of them would make the answer unusable in a medical setting the place seconds matter and belief is earned incrementally.

1. Latency – Clinicians received’t wait. High quality evaluation should return in underneath three seconds to suit naturally into the medical workflow.
2. Accuracy – A number of machine studying (ML) fashions should consider totally different high quality dimensions, together with sharpness, alignment, and completeness, with out false positives that erode clinician belief.
3. Scale – The system should serve tens of hundreds of areas concurrently, with each day volumes within the a whole lot of hundreds.
4. Value effectivity – GPU inference at this scale could be prohibitively costly if not rigorously optimized.
5. International attain – Healthcare is native, however the platform should deploy throughout a number of areas with constant efficiency.

Henry Schein One’s Picture Confirm

Henry Schein One serves dental practices worldwide by way of Dentrix and Dentrix Ascend, follow administration platforms utilized by tens of hundreds of clinicians. Their Platform Companies crew recognized a particular and expensive ache level: as much as 20 % of dental insurance coverage claims are initially denied, with lacking or low-quality photos among the many main causes.

Picture Confirm is an AI-powered high quality evaluation resolution constructed natively into the follow administration workflow. When a technician captures an X-ray, Picture Confirm evaluates it instantly and returns a high quality rating on a 1-to-5 scale. If the picture scores low, the technician retakes it whereas the affected person remains to be current, eradicating the necessity for a return go to.

The product went from idea in fall 2025 to manufacturing in a matter of months. Inside weeks of launch, it was reside in 250 practices. By late April 2026, that quantity had grown to over 10,000, a 43-times improve, with greater than 9 million X-rays processed and weekly volumes averaging 1.5 million and rising.

Picture Confirm is a high quality resolution, not a diagnostic one. It doesn’t establish pathology. It solutions one query: is that this picture ok for medical use? That distinction allowed the crew to iterate with out the regulatory constraints related to medical AI.

For practices, the affect is speedy: fewer affected person callbacks, higher-quality insurance coverage claims, improved coaching for brand new technicians, and a gamification component that drives technician engagement by itself.

“Picture Confirm was solely an concept on the finish of Q3. In 6 months it was created, refined, and now deployed at scale. The design and workflow make adoption quick, intuitive, and scalable.”

— Troy Miller, VP Structure, Henry Schein One

Structure: How AWS and Henry Schein One constructed the answer

Henry Schein One constructed Picture Confirm on AWS from the beginning, utilizing Amazon SageMaker AI for machine studying inference at scale. The applying layer runs on Amazon Elastic Kubernetes Service (Amazon EKS), which orchestrates requests from the follow administration utility to the SageMaker AI inference endpoints and returns the standard rating to the clinician.

Structure diagram:

The inference pipeline

When a picture is captured at a dental follow, it flows by way of a multi-model machine studying inference pipeline hosted on SageMaker AI. The pipeline operates in sequential levels:

1. Classification – The primary stage identifies the picture kind, equivalent to bitewing, panoramic, or periapical, and routes it to the suitable high quality analysis fashions.
2. High quality analysis – Specialised fashions assess sharpness, alignment, protection, and completeness for the recognized picture kind.
3. Rating aggregation – Outcomes from the standard fashions are mixed right into a single 1-to-5 high quality rating returned to the follow administration utility.

The complete spherical journey, from picture seize to high quality rating displayed on display, takes a median of 1.4 seconds with a P90 of two.2 seconds. The system maintains a 0.01 % error price throughout tens of millions of inferences.

The next structure selections enabled the system to fulfill latency, value, and scale necessities:

• SageMaker AI async inference – Handles variable request volumes with out over-provisioning, with autoscaling based mostly on queue depth slightly than CPU utilization, offering a extra correct sign for GPU workloads.
• GPU occasion choice – The crew benchmarked throughout GPU occasion households, in the end migrating from ml.g6e.4xlarge to ml.g7e.4xlarge situations. The newer occasion kind exceeded efficiency expectations. Median latency dropped from 1.687 to 1.432 seconds and P90 from 2.45 to 2.196 seconds. In the meantime, the fleet consolidated from 15 situations all the way down to 10, a 33 % discount in GPU infrastructure with improved response occasions.
• Zero-downtime deployments – An A/B testing framework validates every change in opposition to reside manufacturing site visitors earlier than full rollout, permitting each day optimization iterations with out threat to the manufacturing setting.
• Multi-Area by way of AWS Cloud WAN – Community infrastructure supplies constant world deployment throughout america, Europe, Canada, and Asia Pacific Areas.

The AWS collaboration

When the engagement started, the AWS Options Structure crew reviewed Henry Schein One’s current picture verification workload, which was working on one other cloud supplier. The crew then mapped the structure to equal AWS companies that might ship the identical performance with improved latency and value effectivity at scale. The preliminary focus was on establishing a working baseline on AWS, validating that the multi-model inference pipeline carried out appropriately within the new setting. Because the crew gained confidence within the practical parity, the collaboration shifted towards efficiency optimization and scale. The structure progressively advanced towards a SageMaker AI inference-based resolution that might assist tens of hundreds of areas with sub-3-second response occasions.

All through this journey, the AWS crew offered ongoing structure opinions, helped establish the suitable GPU occasion households for the workload profile, and analyzed utilization patterns to uncover bottlenecks. The crew additionally developed scaling methods aligned with Henry Schein One’s world rollout goal of 40,000 areas. The iterative, hands-on nature of the collaboration allowed each groups to maneuver shortly, delivery optimizations weekly whereas sustaining manufacturing stability by way of zero-downtime deployment patterns.

Optimization: Attaining effectivity at scale

One instructive facet of Picture Confirm’s journey is the optimization story: how the crew recognized infrastructure inefficiencies early, responded shortly, and arrived at a system that right this moment serves tens of hundreds of areas on a lean, extremely environment friendly GPU fleet.

Figuring out the bottleneck

As Picture Confirm scaled quickly in its first weeks of manufacturing, the crew performed infrastructure profiling to grasp utilization patterns. Evaluation revealed that the preprocessing pipeline, together with picture decoding, normalization, and resizing, was working totally on CPU. GPU assets have been being underutilized whereas the general system consumed extra situations than essential to maintain the workload.

This type of CPU-side bottleneck on GPU situations is a typical pitfall in machine studying inference at scale. The sign from CPU saturation can masks GPU headroom, main groups to provision further situations when the actual alternative is pipeline optimization.

The optimization method

The crew recognized a prioritized set of optimization alternatives throughout the pipeline and started delivery them by way of zero-downtime A/B deployments. The primary enchancment, transferring picture preprocessing from CPU to GPU, delivered speedy and important positive aspects in occasion effectivity, with no regression in latency or reliability.

A second optimization adopted shortly after, yielding additional enhancements in throughput per occasion. Inside days, the system was serving a quickly rising location footprint on a fraction of the earlier occasion depend.

“Our crew hung out figuring out 60+ particular issues that might be optimized. We simply began working down the record, typically deploying a number of occasions per day. The three highlights I wish to name out: transferring extra mannequin inference to the GPU immediately (unlocking throughput we are able to’t get CPU-side); altering to an async inference pipeline; and an A/B testing framework, which lets us safely validate enhancements earlier than we push them to 10,000+ areas.”

— Troy Miller, VP Structure, Henry Schein One

Present state

Inside two weeks into the optimization journey, Picture Confirm has processed over 20 million dental X-rays, with weekly volumes averaging 1.5 million and rising. The system serves greater than 10,000 lively areas, a 43x improve from the preliminary 250 practices at launch. It delivers a median latency of 1.4 seconds and a P90 of two.2 seconds with a 0.01 % error price throughout tens of millions of inferences. The fleet runs at roughly 70 % GPU utilization, with 60 % of the optimization backlog already accomplished and enhancements delivery repeatedly.

A key affect is the place these numbers translate to affected person outcomes. As much as 20 % of dental insurance coverage claims are initially denied due to poor picture high quality. Picture Confirm catches low-quality photos on the level of seize, earlier than the affected person leaves, serving to to scale back callbacks, produce cleaner claims, and speed up reimbursement for practices. For clinicians and workplace workers, the suggestions loop is speedy and actionable.

Equally notable is the adoption velocity. The expansion from 250 to over 10,000 areas occurred in weeks, not months. The gamification component, the place technicians see high quality scores for his or her captures, drives engagement with out mandates and creates a pure coaching mechanism for newer workers.

Wanting forward, the structure has been validated for a world goal of 40,000 areas throughout 4 areas. The present 10,000 areas characterize roughly 26 % of that capability, offering important runway for continued development with out re-architecture. The crew treats infrastructure effectivity as a product characteristic, not a one-time undertaking. Delivery optimizations weekly by way of zero-downtime A/B deployments with no buyer affect and no scheduled upkeep home windows.

Classes for machine studying inference workloads

The Picture Confirm optimization journey surfaces 4 rules relevant to machine studying inference workloads working at scale:

1. Profile earlier than scaling – The bottleneck was CPU preprocessing, not GPU compute. Including extra situations would have been costly and ineffective. Instrument the complete pipeline earlier than scaling any part.
2. Optimize the pipeline, not solely the mannequin – Inference latency typically hides in preprocessing, postprocessing, and knowledge motion slightly than in mannequin execution. Profile end-to-end, not solely the mannequin ahead move.
3. Construct for zero-downtime iteration – A/B testing and site visitors shifting assist fast iteration with out manufacturing threat, permitting a each day deployment cadence at scale.
4. Use the suitable autoscaling sign – Queue depth and concurrent request depend are extra dependable scaling indicators than CPU utilization when GPU situations carry combined CPU and GPU workloads.

Wanting forward

Picture Confirm demonstrates a sample relevant far past dental imaging: real-time machine studying inference on the level of care, scaled to hundreds of areas, and optimized to run effectively on minimal infrastructure.

The method rests on 4 key components that organizations can apply to comparable workloads:

1. Managed inference – Amazon SageMaker AI handles the operational complexity of GPU fleet administration, autoscaling, and endpoint lifecycle, liberating engineering groups to deal with mannequin and pipeline high quality.
2. Aggressive pipeline optimization – Profiling the complete inference pipeline, not solely the mannequin, surfaces the actual bottlenecks and delivers effectivity positive aspects with out further infrastructure spend.
3. Zero-downtime deployment patterns – A/B testing and site visitors shifting assist fast iteration with out manufacturing threat, sustaining a each day deployment cadence at scale.
4. Multi-region structure – AWS Cloud WAN and constant infrastructure patterns throughout 4 areas present the worldwide attain that healthcare organizations require.

Henry Schein One’s subsequent section targets 40,000 areas throughout 4 areas, proving that the structure scales not solely technically, however operationally throughout a world healthcare footprint. The crew continues to work by way of its optimization backlog, with every enchancment shipped reside by way of zero-downtime deployments.

For organizations constructing real-time picture verification or comparable machine studying inference workloads, the sample is: begin with managed inference, instrument all the things, optimize the pipeline end-to-end, and construct deployment practices that assist each day iteration.

To study extra about working machine studying inference workloads on AWS, go to Amazon SageMaker AI. To study extra about Picture Confirm and Henry Schein One’s follow administration platform, go to henryscheinone.com.

Concerning the authors